Matt Loper / Brown University Presented for CS296-3 February 14th, 2007 On Three Layer Architectures (Erann Gat) On Three Layer Architectures (Erann Gat)

Matthew Loper Introduction ● What is a good control architecture for a robot? ● How should it coordinate long and short-term thinking? ● Motors and sensors? ● What is the best information flow? ● Example: consider robocup

Matthew Loper Sense, Plan, Act Sense, Plan, Act ● [Nillson, 1980] ● Unidirectional flow of information: nice! ● Why is this bad?

Matthew Loper Sense, Plan, Act Sense, Plan, Act ● [Nillson, 1980] ● Unidirectional flow of information: nice! ● Why is this bad?

Matthew Loper Sense, Plan, Act ● Think about good software design ● Encapsulation: equal amounts of complexity in each box ● Testing: test low-level systems first

Matthew Loper Brook's subsumption idea

Matthew Loper Brook's subsumption idea ● Brook's paper introduced a number of interesting ideas ● Asynchronous communication from different processors in system ● Inhibition of inputs/outputs of lower layers, by upper layers

Matthew Loper Brook's subsumption idea ● A simple robot...

Matthew Loper Brook's subsumption idea ● A wandering robot...

Matthew Loper Brook's subsumption idea ● A wandering, all-seeing, path-planning robot!

Matthew Loper Subsumption: What went wrong?

Matthew Loper Subsumption: What went wrong? ● Not sufficiently modular ● higher-level modules would be intimately involved in details of low-level modules ● Simplistic ● Sometimes, low-level modules should get to interrupt high-level modules ● Not so “taskable”: hand-coded connections may make flexibility low

Matthew Loper 3-Layer Architectures ● In 1991, three groups all came up with basically same approach [Connell91, Gat91, Bonasso91] ● Split the architecture into three modules: ● Deliberative planner ● Sequencer – more on this later ● Reactive feedback controller ● Why three modules? What is the magic here?

Matthew Loper 3-Layer Architectures ● Why three modules? ● Consider different kind of decisions: ● Some require only sensory input, no state ● Some require memory of the past ● Some require both of the above, and planning about the future ● Hence three layers ● Deliberative planner: plans for the future ● Sequencer: remembers the past ● Reactive feedback controller: stateless

Matthew Loper 3-Layer Architectures ● Consider also the separate issue of time/space complexity ● Hence three layers ● Deliberative planner: input is the past and the potential future. May take lots of time in executing long, memory-intensive operations that span many behaviors ● Sequencer: remembers the past, and must be faster ● Reactive feedback controller: should be extremely fast

Matthew Loper Did Freud invent this? ● Freud divided the psyche into “id”, “ego”, and “super-ego” ● Id: inborn, unconscious instincts ● Ego: mediates between the id and super-ego ● Super-ego: intellectualism, morality ● Question: Did Freud invent the 3-layer architecture?

Matthew Loper Did Freud invent this? ● Freud divided the psyche into “id”, “ego”, and “super-ego” ● Id: inborn, unconscious instincts ● Ego: mediates between the id and super-ego ● Super-ego: intellectualism, morality ● Question: Did Freud invent the 3-layer architecture? ● Answer: no, pretty unrelated ● Freud thought everything was about sex

Matthew Loper The 3-layer architecture ● Now let's look at the parts – Reactive Controller – Sequencer – Deliberative

Matthew Loper Reactive Controller ● Lowest level module ● Has programmable “Behaviors”, such as... ● Wall-following ● Collision avoidance ● Moving through doorways ● This module should be ● fast, low/memory ● able to detect failures and deal with them ● without internal state

Matthew Loper Sequencer ● Select primitive Behavior, and supply parameters to that Behavior ● Has some state about the past ● Performs tasks that do not require planning

Matthew Loper Deliberator ● Can either send tasks to sequencer, or just respond to sequencer queries, depending on application ● Example problem spaces: ● Localization ● Path Planning ● Motor Planning

Matthew Loper The Debate: Planning vs. Reacting ● What's the debate again? ● “Some of this passion may be the result of a hidden conviction on the part of AI researchers that at the root of intelligence lies a single, simple, elegant mechanism. But if, as seems likely, there is no One True Architecture, and intelligence relies on a hodgepodge of techniques [my emphasis], then the three-layer architecture offers itself as a way to help organize the mess”

Matthew Loper What this paper doesn't cover ● The paper doesn't cover a number of important issues: ● Sensor processing ● Learning ● World modeling ● But its take-home message is good; ● Algorithms seems to fall into three groups based on time-dependency (stateless, sequencing, or deliberative) ● Still build and test from the bottom up

Matthew Loper References ● Eran Gat. “On Three-Layer Architectures.” Artificial Intelligence and Mobile Robotics, in D. Kortenkamp, R. P. Bonnasso and R. Murphy (eds.), AAAI Press, pages , ● Rodney A. Brooks. “A Robust Layered Control System for a Mobile Robot.” IEEE Transactions on Robotics and Automation, 2(1), pages 14-23, April ● Stole diagram from: –